Rigor and flexibility: the benefits of agent-based computational economics

The development of agent-based computational economics (ACE) has been driven by two main goals: rigor and flexibility. ACE combines the principles of economics and mathematics with powerful high performance computer simulation to develop agent-based models for the economy. This combination gives ACE a unique capacity to help policy makers design effective solutions to complex economic problems.

In this blog post, we will provide an overview of ACE, its history, and how it can be used to benefit policy design.

What is ACE?

The primary goal of many agent-based modelers is to create large-scale simulations that represent the behavior of complex social and economic systems. ACE specializes this goal to economic systems. It has become increasingly popular over the past two decades, with researchers attempting to use it to provide insights into complex phenomena like income inequality, the impact of climate events on the economy or the tragedy of the commons.

Agent-Based Computational Economics (ACE) focuses on agent decision making and stochasticity – two key components of economic modeling. Agent decision making describes how each agent in a given system makes choices based on their goals and perceived environment, while stochasticity describes the unpredictable or probabilistic nature of such decisions. ACE research objectives also focus on creating models that accurately simulate the interactions between agents, as well as the outcomes of their collective behavior.

Agent behavior is simulated through computer algorithms and may be defined by a set of rules and parameters that reflect the behavior of real people in a given economic context. NetLogo is an open source platform for developing agent-based models. It offers a wide range of capabilities for simulating the behavior of agents. Machine learning may also be used to automatically discover agent behaviors.

The development of ACE can be traced back to the work of W. Brian Arthur, father of Complexity Economics who argued for the importance of heterogeneity in economic behavior and dynamics. His work was later extended by economists Leigh Tesfatsion¹, Rob Axtell and many others, who used ACE to better understand how different economic processes interact.

“ACE is an economics in which the agents in the economy are realistically human and realistically diverse, in which path-dependence and history matter, in which events trigger events, and in which the networks that channel these events matter. It is an economics in which equilibrium is not assumed, if it’s present it emerges; in which rational behavior is not assumed, in general it is not well-defined; in which the unexpected crises of the economy can be probed and planned for in advance; in which free markets are not assumed to be optimal for society but can be assessed realistically; and in which distributional issues are not covered up, but can be rigorously scrutinized.”

Brian Arthur – Foundations of Complexity Economics – Nature

Today, ACE is used in many different fields, from game theory to macroeconomics, as it provides a powerful tool for understanding complex economic systems. By studying the behavior of agents in computer simulations, researchers can gain insights into the underlying dynamics of economic systems, as well as their implications for policy design.

Overall, this is a powerful tool for economic policy design and decision making. By creating large-scale simulations that represent the behavior of complex economic systems, researchers can gain insight into how different policies may affect market outcomes and the welfare state of society at large.

A Brief History of ACE

The development of Agent-based Computational Economics (ACE) was motivated by the need to bring together flexibility and logical rigor. ACE was developed in the 1990s by Chen and Yeh, who created an agent-based stock market model that included stock market traders as well as a “business school.” This business faculty used test results to revise their models. Furthermore, the traders evolved their forecasting models through individual and social learning. In 1993, Leigh Tesfatsion published “Evolution of Strategies for Multiagent Environments” as part of the ISU Alife workshop, which also had an important impact in the literature of the subject.

In 2006, the ACE handbook volume was published, and since then the research areas making use of agent-based modeling for the study of economic systems has been rapidly expanding. These research areas include coupled economic and ecological systems; experiments with real and computational agents; macroeconomics; circular economy and technological innovation.

Researchers often use large-scale simulations to explore the complex interactions between agents and economic systems. This means we can investigate how different types of policies interact and affect each other. It makes it possible to examine the interaction between macroeconomic policies and microeconomic behavior over time.

Large-scale simulations can also be useful in studying how different policies interact at both the national and international levels – like investigating how trade policies might impact foreign investment decisions, and therefore income inequality. We can also use them to examine how different types of policies interact with one another over time – like how taxation policies might affect different types of income earners, or how the minimum wage affects businesses’ hiring decisions.

By simulating policies, researchers can gain insight into how different factors interact on a macro level and can better inform policy design, like simulating the effects of a Universal Basic Income on overall economic growth, income distribution and welfare state of a population.

“The development of modern technologies and computing capabilities has created new opportunities to test the impact of public policies on the economy, financial system and society (social welfare). So far, central banks have primarily used Dynamic Stochastic General Equilibrium models to assess the effects of a selected monetary, financial or macroprudential policies. Nowadays, large scale data-driven agent-based models are a tempting alternative to these models. Due to the greater flexibility and realistic assumptions of these models, the agent-based approach is gaining in importance in central banking and financial supervision. Because of the cooperation with AWS, it was possible to run an agent-based simulation for Poland and to prepare the infrastructure to simulate the effects of policies for many EU countries simultaneously under the MACROPRUDENTIAL project. The primary aim of this project³ was to investigate how new macroprudential policies can influence financial stability without contributing to inequality in society. In this project, we applied cutting-edge, agent-based simulation techniques to uncover the redistributive effects of macroprudential policies and to examine the optimal combination of macroprudential tools from a social welfare perspective. The results of this project complement the conclusions that have been drawn from the ECB’s system-wide stress-testing exercises by providing data on the rise of inequality in selected EU countries after it adopted new financial regulations. Processing such a large amount of data and the complexity of the calculations in the simulation required the use of the services offered by AWS. The support and guidance they provided was invaluable and largely contributed to the success of the entire project”.

Jagoda Kaszowska-Mojsa, Oxford INET/ Bank of Poland

Reinforcement Learning and ACE

Reinforcement Learning (RL) is a type of Artificial Intelligence (AI) that is becoming increasingly important in the field of Agent-Based Computational Economics (ACE). RL can be used to model the agents in an economic system, allowing them to learn from their environment and refine their decisions by assessing their successes or failures. Through this process, an agent can become better and better at making decisions and more rational in its simulated behavior.

In the context of ACE, reinforcement learning is also useful for determining the optimal economic policy for a certain environment. Here, an AI agent can explore the different economic policies available, and use the feedback from the environment (which may also include AI agents to simulate participants) to evaluate the outcomes of a policy. This way, it can identify which policies are more successful than others and which policies we should avoid.

The use of reinforcement learning has been particularly useful in developing the AI Economist², which is an AI system capable of creating its own economic policies. By using reinforcement learning, the AI economist can develop economic policies based on its experience and gain a better understanding of how to best promote economic well-being. This allows the AI Economist to create policies tailored to the needs of a specific environment.

“A new capability in agent-based modeling is the use of multi-agent reinforcement learning to simulate the behavior of participants in an economy. This allows the modeler to focus on correctly capturing the motivations and information available to different kinds of participants, rather than anticipating the behaviors of participants. Rather, the behaviors can then be learned through a very large number of simulated interactions in the economic environment.”

Dr. David Parkes, AI Economist co-author, Harvard Data Science Initiative

The Benefits of ACE

Agent-based Computational Economics (ACE) provides rigor and flexibility when modeling and predicting economic outcomes, due to its ability to capture emergent phenomena. By employing a natural description of the system as opposed to relying on assumptions or approximations, ACE is able to account for dynamic interactions between individuals, organizations, and institutions that traditional economic models may miss. This allows ACE to provide accurate predictions and helps us better understand the behavior of economic systems.

It also offers an unprecedented level of flexibility. By simulating multiple scenarios and providing instant feedback on the consequences of changes in policy and market conditions, ACE can explore a wide range of possible outcomes and anticipate unexpected behaviors. This makes it an invaluable tool for policy makers and climate-risk managers who are looking for new ways to stress test macro scenarios or the precise impact of climate events on geo-located assets.

The ability of ACE to deal with emergent phenomena is also what drives its other benefits. For example, ACE can process large amounts of data quickly thanks to its ability to leverage HPC. This makes it well suited for problems where high-speed and great accuracy are essential, like in financial and climate risk management, or epidemiology – enabling organizations to deploy powerful predictive models, simulate complex scenarios, and identify potential risks before they become reality.

Doyne Farmer’s group at Oxford’s INET (Institute for New Economic Thinking) made an agent-based model of the economic effect of the COVID pandemic on the United Kingdom and used it to make real-time predictions. They tested the economy impact of several possible lockdown policies, and predicted that, if the government adopted the policy they recommended, the drop in GDP in the second quarter of 2020 would be 21.5%. The government did adopt the recommended policy, and the actual economic impact was 22.1%. Diving deeper, they made reasonably good predictions for most sectors of the economy, and for the behavior through time. This provides a proof-of-principle demonstrating that agent-based models can be used for prediction and policy evaluation.

Running ACE on AWS HPC Clusters

The computational power required to run Agent-Based Computational Economics (ACE) call for the help of an HPC cluster. By leveraging HPC, researchers can more quickly and efficiently carry out their experiments and simulations.

One way to access the necessary resources for an HPC cluster is through Amazon Web Services (AWS). AWS offers parallel clusters on its Elastic Compute Cloud (EC2) platform that enable researchers to create virtual servers on-demand for running ACE. These EC2 clusters allow for the rapid deployment of ACE simulations and experiments with minimal cost and effort.

We measure the costs of running an ACE simulation on an Amazon EC2 by counting instance hours, or the number of hours an instance is running. An Amazon EC2 cluster can provide cost-effective performance that can help researchers speed up their ACE simulations.

AWS also provides a variety of tools and services to help manage, scale, and monitor ACE simulations on these clusters. For example, AWS allows users to set up automated backups, track data, and compute resources, and set up multiple simulations on a distributed fashion. All these features can help researchers minimize time spent on managing ACE simulations on their EC2 clusters.

Overall, the ability to run ACE simulations on AWS HPC clusters allows researchers to take advantage of a powerful platform to explore and analyze their data quickly and efficiently. The combination of AWS’s cost-effective performance and its range of management tools makes it an ideal platform for running ACE simulations.

Conclusion

The flexibility, logical rigor, and ability to capture emergent phenomena that ACE offers make it an invaluable tool for policy makers and researchers alike. Its combination of speed and accuracy, coupled with its capacity for dealing with large datasets, make it a powerful tool that can provide critical insights into our understanding of economic systems.

If you are interested in discussing how to develop you own ACEs, we would like to hear from you. Reach out to us at ask-hpc@amazon.com.

References

1. Leigh Tesfatsion
2. AI Economist
3. This project has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 101023445 and was developed at the University of Oxford.